Datasets:

sochastic
/

ChemPro

pretty_name: ChemPro
license: cc-by-nc-4.0
language:
  - en
size_categories:
  - 1K<n<10K
task_categories:
  - question-answering
  - multiple-choice
task_ids:
  - multiple-choice-qa
tags:
  - chemistry
  - science
  - benchmark
  - evaluation
  - reasoning
  - curriculum
  - arxiv:2602.03108
configs:
  - config_name: mcq
    data_files:
      - split: easy
        path: data/mcq/easy.parquet
      - split: medium
        path: data/mcq/medium.parquet
      - split: challenging
        path: data/mcq/challenging.parquet
      - split: difficult
        path: data/mcq/difficult.parquet
  - config_name: numerical
    data_files:
      - split: easy
        path: data/numerical/easy.parquet
      - split: medium
        path: data/numerical/medium.parquet
      - split: challenging
        path: data/numerical/challenging.parquet
      - split: difficult
        path: data/numerical/difficult.parquet

ChemPro

A progressive chemistry benchmark for large language models.

ChemPro is an evaluation benchmark. It contains 4,100 chemistry questions in four tiers of increasing difficulty.

Most benchmarks take their difficulty labels from annotator judgement. ChemPro does not. Each tier comes from one established educational source:

elementary web material
Indian NCERT curricula for grades 9–10 and 11–12
JEE Mains competitive examinations

Decades of curriculum design set the difficulty of each tier. No one applied the labels afterwards.

The benchmark keeps conceptual understanding separate from computational reasoning. Multiple-choice questions test the first. Numerical questions test the second. You can score the two independently.

Why the tiers matter

Many models score well on introductory chemistry. Their accuracy then drops sharply as the questions become harder to read, even when the concepts stay inside the same syllabus.

JEE Mains follows the official NCERT syllabus. The challenging and difficult tiers therefore cover the same concepts. The accuracy gap between these two tiers shows the effect of question complexity alone, not of subject coverage.

Split	`tier`	Source	MCQ	Numerical	Total
`easy`	`CP_E`	Web quizzes and questionnaires	590	204	794
`medium`	`CP_M`	NCERT, grades 9–10	229	107	336
`challenging`	`CP_C`	NCERT, grades 11–12	455	208	663
`difficult`	`CP_D`	JEE Mains, 2020–2024	1,493	814	2,307
Total			2,767	1,333	4,100

Each question also has one or more subfield labels:

Subfield	Items
Inorganic-Chemistry	1,651
Physical-Chemistry	1,536
Organic-Chemistry	915
Bio-Chemistry	227

Usage

from datasets import load_dataset

# One tier of multiple-choice questions
mcq = load_dataset("sochastic/ChemPro", "mcq", split="difficult")
print(mcq[0]["question"])
print(mcq[0]["choices"])
print(mcq[0]["answer"], mcq[0]["answer_text"])

# All tiers at once
every_tier = load_dataset("sochastic/ChemPro", "mcq")

# Numerical problems
num = load_dataset("sochastic/ChemPro", "numerical", split="easy")
print(num[0]["question"], num[0]["answer_value"], num[0]["answer_unit"])

To filter by subfield:

organic = mcq.filter(lambda r: "Organic-Chemistry" in r["attributes"])

Schema

Both configs share these fields:

Field	Type	Description
`id`	string	A stable identifier, for example `chempro-difficult-mcq-01504`. You can cite it.
`source_id`	int32	The index of the question in its source tier and question type
`tier`	string	The tier symbol used in the paper: `CP_E`, `CP_M`, `CP_C` or `CP_D`. The split name gives the readable form.
`source`	string	`web`, `ncert_grade_9_10`, `ncert_grade_11_12` or `jee_mains_2020_2024`
`question_type`	string	`mcq` or `numerical`
`question`	string	The question stem. For MCQ items, the options are in `choices`.
`attributes`	list[string]	One or more of the four subfields above

The mcq config adds these fields:

Field	Type	Description
`choices`	list[string]	Exactly four options, in the order A to D
`answer`	string	The letter of the correct option, `A` to `D`
`answer_index`	int32	The zero-based index of the correct option in `choices`
`answer_text`	string	The text of the correct option. It always equals `choices[answer_index]`.

The numerical config adds these fields:

Field	Type	Description
`answer`	string	The answer as written. It includes the unit when the source gave one.
`answer_value`	float64	The numeric value taken from `answer`. Use it for tolerance-based scoring.
`answer_unit`	string	The unit taken from `answer`. It is null when the answer has no unit.

Examples

An MCQ item:

{
  "id": "chempro-difficult-mcq-00001",
  "tier": "CP_D",
  "source": "jee_mains_2020_2024",
  "question_type": "mcq",
  "question": "The five successive ionization enthalpies of an element are ...",
  "choices": ["2", "4", "3", "5"],
  "answer": "C",
  "answer_index": 2,
  "answer_text": "3",
  "attributes": ["Inorganic-Chemistry"]
}

A numerical item:

{
  "id": "chempro-easy-numerical-00001",
  "tier": "CP_E",
  "question_type": "numerical",
  "question": "Calculate the molar mass of $$ CH_3COOH $$.",
  "answer": "60.05 g/mol",
  "answer_value": 60.05,
  "answer_unit": "g/mol",
  "attributes": ["Physical-Chemistry"]
}

Chemical notation stays in LaTeX between $$ delimiters. Remove the delimiters or render them, as your evaluation requires.

Evaluation notes

MCQ. Score against answer, which gives the letter, or against answer_text. The options are an ordered list, so you can shuffle them to control position bias. If you shuffle them, set answer_index again.
Numerical. answer_value supports exact-match scoring and tolerance-based scoring. Tolerance-based scoring tells you more, because many items round at intermediate steps. The unit is a separate field, so a correct value does not lose marks for its format.
Report the result for each tier. A single average hides the fall in accuracy that this benchmark shows.

Construction

We took the questions from the sources above. We put them into one text format. Three passes then verified them:

A source check confirmed the origin of each question.
An expert review confirmed that each question is correct and has one meaning.
Automated checks looked for format errors and duplicates.

GPT-4o applied the subfield attributes. Human annotators did not. Use these labels to slice the data. Do not use them as a gold-standard taxonomy.

Each tier holds its full number of questions. Every question matches the difficulty level and the subject mix of its tier.

Limitations

Possible exposure. The difficult tier comes from the JEE Mains papers of 2020 to 2024. These papers are public and may appear in pretraining data. The paper estimates roughly 8% possible exposure. It used four probes: prefix completion, paraphrase detection, content modification, and reverse engineering. A comparison with other benchmarks found minimal overlap with existing chemistry datasets.
Text only. Some questions first used chemical structure diagrams. These questions now use text. We rewrote or removed the items that text could not show correctly. The benchmark therefore under-represents structure recognition.
Curricular scope. The difficulty comes from the Indian secondary and competitive examination system. It applies well to general chemistry ability. Do not read the tier names as universal difficulty labels.
Answer keys. Every item passes automated structural validation. We cross-checked the answer keys wherever two near-identical questions disagreed. This method cannot find an error that both copies of a question share. It also cannot find an error in an item that has no near-duplicate. Assume a small number of remaining errors, as in any benchmark of this size.
Language. The dataset is in English only.
No training split. ChemPro is an evaluation benchmark. The tiers are splits for convenience. No split is intended for training.

Licensing and provenance

ChemPro uses the CC BY-NC 4.0 licence. You can share and adapt the dataset for non-commercial purposes. You must give attribution.

The medium and challenging tiers come from NCERT curricular material. The difficult tier comes from JEE Mains examination papers. The National Testing Agency administers those examinations. We adapted the questions. We did not copy them word for word. The non-commercial term follows from these sources. If you plan to use the dataset for more than non-commercial research, check the status of the source material first.

Citation

If you use ChemPro, please cite the journal article:

@article{baranwal2026chempro,
  title   = {ChemPro: A progressive chemistry benchmark for Large Language Models},
  author  = {Baranwal, Aaditya and Vyas, Shruti},
  journal = {Artificial Intelligence Chemistry},
  volume  = {4},
  number  = {1},
  pages   = {100118},
  year    = {2026},
  issn    = {2949-7477},
  doi     = {10.1016/j.aichem.2026.100118},
  url     = {https://www.sciencedirect.com/science/article/pii/S2949747726000126}
}

The preprint:

@article{baranwal2026chempro_arxiv,
  title         = {ChemPro: A Progressive Chemistry Benchmark for Large Language Models},
  author        = {Baranwal, Aaditya and Vyas, Shruti},
  journal       = {arXiv preprint arXiv:2602.03108},
  year          = {2026},
  doi           = {10.48550/arXiv.2602.03108},
  url           = {https://arxiv.org/abs/2602.03108}
}

Paper: https://www.sciencedirect.com/science/article/pii/S2949747726000126
Preprint: https://arxiv.org/abs/2602.03108

Contact

Aaditya Baranwal, University of Central Florida. Email: aaditya.baranwal@ucf.edu